Seeing what @wiz_io achieved with Gemini 4 Argon is amazing – and it makes it all the more frustrating that on the consumer/developer side, Argon is still so heavily restricted by safety filters that we can barely use it for cybersecurity ! 😅🤍 I do authorized vulnerability research – mostly on Google’s own services through Google VRP, AI VRP, and OSS VRP – and I’m already a Trusted Tester on both @antigravity and @GeminiApp. Yet whenever I try to use Argon for legitimate security research, it is constantly throttled by : « This request was blocked by Gemini's filters. They can occasionally trigger by mistake on safe coding, security, or biology-related queries. » I’ve already applied to the Fairwind Program (screenshot below), but I still haven't received an invitation. Could someone at Google help review my Fairwind application so independent VRP researchers helping secure Google’s own ecosystem can also use Argon without these guardrails ? 🙏🤍✨ cc: @Google, @GoogleDeepMind, @GoogleVRP, @antigravity, @joshwoodward, @OfficialLoganK, @rodydavis, @_mohansolo

Actually, I see what Qwinah (who Rody is replying to) means, but it is poorly phrased – Argon does not hallucinate on historical or important facts, it is much more specific and localized than that. (Rody, this one is for you too ! And @GeminiApp team & @joshwoodward, read this please – it explains where the issue comes from, hoping you’ll fix these problems ! :D) As a French user who uses Gemini for pretty much everything, I notice that Gemini, even on Argon, has a tendency to add unnecessary details to make the sentence feel more complete, and that is what causes the model to introduce hallucinations into its response. The core of the responses is often very good, but in order to make its answer feel more complete, the model adds extra details – often false or exaggerated – to improve the fluidity and impact of its points. And on niche or poorly documented topics, especially as a French user, I see this a lot haha ! The model will know the basics of a topic, then try to talk about details it doesn't actually know, and end up making a ton of mistakes ! (Often, it has a reflex of adding details inside "()" in its sentences, and that is where a huge amount of hallucinations can be found !) And the Gemini App doesn't help, having a terrible system prompt with an intent router system for almost every tool, forcing the model EITHER to search the web OR not to search AT ALL for the message, which makes the responses terrible because each moment needs a personalized way of handling the response, and this intent router is not the solution for that. And a little note on Search : web search has no persistence in Gemini App chats, so the model cannot stay coherent across the conversation because previous results found by the web search tool are not saved in the chat context ! 😅 Gemini Spark has that partially – it retains the first 3 results (out of 45, with the other 42 stored in the sandbox if it needs to inspect them, though keeping 10 at least directly in context would be much better ;-;), so at least we have some sort of "retention" there (Antigravity has tool persistence too) – but you really should fix that : when users are talking about a very recent or fast-moving topic, without near-live context persisted across turns, the model cannot stay precise in its sentences without getting things wrong, which is so disappointing ! Tool persistence needs to be handled in the Gemini App. 🙏🤍 The YouTube tool guidelines, which trigger contextually, make it recommend videos to users that are sometimes poorly suited to the context – Gemini is forced to do it, and it completely breaks the flow of the chats. (Or even the obligation to describe videos with timestamps and a very "institutional" way of describing videos because of that, this is very annoying !) Since I use Antigravity a lot now, I see it less, but many people using Gemini through the Gemini App have a completely broken experience because of a harness that poorly manages how Gemini can or cannot take actions. Agency was very poorly designed by the @GeminiApp team : instead of letting Gemini CHOOSE when to do WHAT rather than placing intent routers everywhere, they put intent systems all over the place. The model should have all of its tools defined right from the start of the chat – code execution enabled, web access available so it can search WHENEVER IT WANTS, URL Context active, personal context of course, and less rigid, contradictory instructions that paradoxically make Gemini worse in chat, and sometimes even in coding : since code execution only triggers when a very specific data-science or professional need is detected (analyzing a .csv file for instance), it makes Gemini unusable for coding, as Gemini cannot launch its sandbox to verify the accuracy of its code. And in Notebooks, code execution cannot be launched, even when the intent router accepts code execution, because there is a backend bug that I've already reported in a previous post, that has still not been fixed ! Gemini doesn't even have Node in its sandbox either, which is such a shame ! And it cannot download GitHub repos either, whereas even ChatGPT can do it via a GitHub tool where it can also perform GitHub actions, while Gemini is strictly read-only ! (And the GitHub tool is only available for Gemini when giving an official link of our repo on the Gemini App "import code" option) There is a real issue with how instructions are managed internally in the Gemini App, and with its overall agentic harness. Thankfully, Antigravity fixes a lot of these issues, but this ruins the average user experience on the Gemini App and creates a biased perception of Gemini among the general public, even though the model itself is incredible ! 🥹❤️ cc: @joshwoodward, @tulseedoshi, @rodydavis, @OfficialLoganK, @GeminiApp, @GoogleDeepMind

It was very fast 🙏🤍✨ Thanks Google, you are the best ! 🥹❤️

I actually signed up on this exact form back in the summer, and like many people in the replies to Josh’s post, I only see the "You've already responded" screen without having received an onboarding email yet 😅 I just updated my response on the form now that I’m 18 (I wasn't when I first filled it out, so it could be because of that) – and after having tested Gemini 4 Argon in early access as a Trusted Tester on @antigravity (and sharing deep technical breakdowns on the @GeminiApp architecture in my public posts), I would genuinely love to be part of the @GeminiApp Trusted Tester community too ! 🙏🤍✨ cc: @joshwoodward, @GeminiApp

My Gemini is chill compair to lot of users ! 🙈🤍